Reduced rich-club connectivity is related to disability in primary progressive MS.

OBJECTIVE: To investigate whether the structural connectivity of the brain's rich-club organization is altered in patients with primary progressive MS and whether such changes to this fundamental network feature are associated with disability measures.
METHODS: We recruited 37 patients with primary progressive MS and 21 healthy controls for an observational cohort study. Structural connectomes were reconstructed based on diffusion-weighted imaging data using probabilistic tractography and analyzed with graph theory.
RESULTS: We observed the same topological organization of brain networks in patients and controls. Consistent with the originally defined rich-club regions, we identified superior frontal, precuneus, superior parietal, and insular cortex in both hemispheres as rich-club nodes. Connectivity within the rich club was significantly reduced in patients with MS (p = 0.039). The extent of reduced rich-club connectivity correlated with clinical measurements of mobility (Kendall rank correlation coefficient τ = -0.20, p = 0.047), hand function (τ = -0.26, p = 0.014), and information processing speed (τ = -0.20, p = 0.049).
CONCLUSIONS: In patients with primary progressive MS, the fundamental organization of the structural connectome in rich-club and peripheral nodes was preserved and did not differ from healthy controls. The proportion of rich-club connections was altered and correlated with disability measures. Thus, the rich-club organization of the brain may be a promising network phenotype for understanding the patterns and mechanisms of neurodegeneration in MS.

MS is the most common autoimmune disease of the CNS, and persistent inflammation as well as chronic progressive neurodegeneration in the brain and spinal cord leads to accumulation of disability.[1,2] MRI is currently the best available surrogate marker of MS pathology.[3] MRI also permits the use of probabilistic tractography, which allows investigating the integrity of structural connections in the brain.

The human brain can be considered as a network and the network's topology can be studied by graph theory. The network perspective may offer new insights into disease-specific processes such as neurodegeneration and has been applied to several neuropsychiatric diseases.[4,5] Recent studies recognized an essential topological feature of the human connectome. The brain network is organized into a so-called rich-club and peripheral nodes.[6,7] The rich-club brain regions are more densely interconnected than expected by chance and form a prominent subnetwork of the brain. They are considered to have a guiding function controlling integration and information flow in the brain network. The rich-club architecture assures a highly efficient structural and functional network organization,[8] but might be vulnerable in brain diseases. Altered rich-club connectivity has been observed in schizophrenia,[9] migraine,[10] or dementia.[11]

MS can also be considered as a network disorder leading to focal and global impairment of the brain network.[5,12-16] However, the rich-club organization in patients with MS has not been analyzed. We aimed to analyze the network topology of structural connectomes of patients with a predominantly neurodegenerative disease course. Primary progressive MS (PPMS) shows less and more diffuse inflammatory disease activity than relapsing-remitting MS (RRMS), and clinical assessment in most cases is not blurred by superimposed relapses. Therefore, PPMS might be taken as a candidate model for investigating long-term changes of brain network architecture in MS. We hypothesized that the connectome architecture of patients with PPMS differs from that of healthy controls (HCs), in particular with respect to the rich-club organization, and that these changes might be associated with disability measures.

METHODS

Patients and controls.

Patients were eligible for this observational cohort study if they were diagnosed with PPMS according to the McDonald criteria 2010[17] and had an Expanded Disability Status Scale (EDSS) score of ≤7.0. Patients (n = 37) obtained structural MRI and completed a modified multiple sclerosis functional composite (MSFC) test battery,[18] including the timed 25-foot walk (T25FW, short distance walking speed), the Nine-Hole Peg Test (NHPT, dominant and nondominant fine motor hand function), and the Symbol Digit Modalities Test (SDMT, information processing), which is considered as a simple-to-administer, less stressful, and valid substitute for the MSFC standard Paced Auditory Serial Addition Test.[19] We recruited 21 HCs matched by age and sex and applied the same MRI protocol.

Standard protocol approvals, registrations, and patient consents.

Participants were recruited at our MS day hospital (2012–2016), provided written informed consent, and the Institutional Review Board (ethics committee of the Hamburg Chamber of physicians, PV3961/PV4405) approved the study.

Image processing and reconstruction of brain networks.

Briefly, the MRI protocol included a T1-weighted sequence (0.9 × 0.9 × 0.9 mm), a T2 sequence (0.5 × 0.5 × 3.0 mm), and diffusion tensor imaging (single-shell, 20 directions with noncollinear diffusion gradients [b = 1,000 s/mm2] and 1 nondiffusion-weighted b0 image, 1.9 × 1.9 × 2.0 mm). Images were processed with the functional imaging software library (FSL) and FreeSurfer software.[20] For each participant, the gray matter was parcellated into 34 cortical regions per hemisphere and 8 subcortical regions based on the Destrieux atlas. Subcortical white matter regions corresponding to the 34 cortical regions were used for FSL probabilistic tracking with crossing fibers (probtrackx). Based on the number of streamlines reaching from one FreeSurfer region to another, 2 different kinds of networks were defined. First, we used the average number of streamlines (forward/backward) between every 2 regions as edge weight to construct weighted networks Graw. Second, binary networks (0 = unconnected and 1 = connected) were constructed (Gbin) based on Graw, by discarding connections with a number of streamlines below a given threshold (e.g., 80% of maximum connection strength). This procedure meets the standard to avoid totally connected networks and reduces the number of edges to the most prominent and strongest connections. However, there is no consensus, how to define this threshold.[5,21] We investigated the small-world index (SWI) of individual Gbin at thresholds from 2.5% to 97.5% in steps of 2.5% to determine a suitable common cutoff for all participants. The best threshold was defined as the highest cutoff with a median SWI above 1 (indicating small-world features) but low variance of the SWI (no major bias of this fundamental network feature). For a detailed description, see supplemental material at Neurology.org/nn.

Graph metrics.

Global parameters included strength (Graw), average shortest path length (APL, Gbin), global efficiency (Gbin), clustering coefficient (Gbin) and weighted clustering coefficient (Graw), and arithmetic mean method. In addition, we computed the node-specific degree (Gbin), betweenness centrality (Gbin), and strength (Graw).

Rich club.

Rich-club nodes are more densely interconnected than expected by chance (compared with random networks). Thus, the rich club of a given network can be defined by top-ranking nodes based on degree or strength. Eight cortical regions have been identified as rich-club hubs in healthy individuals, which we used as an a priori–defined rich club: superior frontal, precuneus, superior parietal, and insular cortex in both hemispheres.[6] First, we were interested if node-specific measures identify the predefined rich-club nodes in terms of high strength, degree, and betweenness in patients and controls. We then implemented a formal test of the rich-club organization[22] and investigated whether alternative rich-club definitions (accounting, e.g., for individual variability) might perform better than the original definition (see supplemental material). We computed the connectivity within the rich-club, between rich-club and peripheral nodes (so-called feeder connections), and between peripheral nodes. In addition to the absolute values, rich club, feeder, and peripheral connectivity were each divided by the total connectivity of the individual network. The latter approach was chosen to account for the presumed generalized loss of connectivity in patients with PPMS. The measures were then compared by their ability to distinguish between patients with PPMS and HCs, and their association with clinical outcomes and global MRI volumes.

Statistics.

We performed descriptive statistics according to the nature of the data as mean with SD or as frequencies and/or percentages. Differences between patients and controls were assessed by the Student t test for continuous data and the χ2 test for categorical data. The distribution of connectivity on rich club, feeder, and peripheral connections was compared by multivariate analysis of covariance (MANCOVA adjusting for age, sex, T2 lesion volume, and total connectivity). To depict the direction of changes, relative connectivity in the different compartments was further analyzed by t tests. To account for nonnormal distribution of data, the association between variables was investigated nonparametrically by computing 1-sided Kendall correlation coefficient τ. Post hoc, we compared the predictive value of rich-club connectivity for clinical outcomes with multivariate linear regression. We used analysis of variance to compare unadjusted models with models adjusted for brain volume and T2 lesion volume. p Values below 0.05 were considered statistically significant. False discovery rate (FDR) was used to correct for multiple testing. All analyses were performed with Statistics in R 3.2.3, including the igraph and tnet packages.[23,24]

RESULTS

Cohort and threshold selection.

Table 1 summarizes descriptive statistics. The mean age of patients with PPMS was 52.2 years and did not differ from HCs (50.4, p = 0.282). Disease duration since first symptoms was 7.7 years, and a median EDSS of 3.5 (range 1.5–7) indicated a moderate disability. Brain volume was lower in patients (p = 0.003).

Table 1

Descriptive statistics

Stepwise thresholding of Graw networks increased the SWI of networks only if higher cutoffs were applied (figure e-1). Only with a threshold above 95%, the median SWI was above borderline values. However, we observed in cutoffs above 82.5%, an increase of the variability and several outliers in patients and controls. We interpreted this finding as artificial noise in the data and, therefore, decided to construct Gbin based on the highest threshold with acceptable variability, which was 82.5%.

Global graph metrics.

Global network metrics are presented in table 2. There was no difference between patients with PPMS and controls in terms of total connectivity (p = 0.340) or any other graph metric. The association between global graph metrics, MRI volumes, and clinical data is summarized in table 3. Although age, T1, and T2 lesion volumes did not correlate with graph metrics, strength decreased with longer disease duration (τ = −0.30, p = 0.005), whereas APL increased (τ = 0.23, p = 0.027) with longer disease duration. APL correlated inversely with white matter volume (τ = −0.21, p = 0.032) and global efficiency decreased accordingly (τ = 0.23, p = 0.025). Walking speed (T25FW) was associated with strength (τ = −0.28, p = 0.009) and APL (r = 0.28, p = 0.011). The association between strength and disease duration, respectively, T25FW, and the correlation between APL and T25FW remained after FDR correction.

Table 2

Graph metrics: patients and controls

Table 3

Association of global graph metrics, MRI volumes, and clinical data in PPMS

A priori–defined rich-club nodes showed the highest betweenness of all nodes (figure 1), which indicates that their role as an important junction within the networks was preserved. These nodes also ranked within the top 9 nodes based on strength and degree (figures e-2 and e-3). Overall, the large-scale organization of the connectomes did not differ between patients with PPMS and HCs. The order of nodes was highly correlated between patients and controls, based on strength (r = 0.99, p = 0.001), degree (r = 0.99, p < 0.001), and betweenness (r = 0.92, p < 0.001). Formal testing confirmed the rich-club organization in average connectomes of patients and controls (figure e-4).

Figure 1

Node-specific graph metrics: Betweenness

Boxplots ordered by median values of nodes of Gbin. (A) Healthy controls and (B) primary progressive MS.

For the a priori–defined rich club, absolute connectivity strength was lower in patients with PPMS than that in controls (p = 0.038, FDR-corrected p = 0.114). The absolute strength of feeder and peripheral connectivity did not differ (p = 0.317 and p = 0.474, FDR-corrected both p = 0.474). Corrected for total connectivity, T2 lesion volume, age, and sex, we observed an altered distribution of connectivity between rich club, feeder, and periphery in patients with PPMS (MANCOVA p = 0.011): pairwise T tests revealed that the percentage of connections within the rich-club, that is, the relative rich-club connectivity, was reduced in patients (p = 0.013, FDR-corrected p = 0.039), whereas peripheral connectivity was relatively increased in patients than that in controls (p = 0.040, FDR-corrected p = 0.060, figure 2, A–C).

Figure 2

Rich-club connectivity

Schematic representation of the rich-club organization in average primary progressive MS (PPMS) connectomes (top right). Boxplots show differences between PPMS (blue) and healthy controls (HCs) (orange) in relative connectivity within the rich club (A), for feeders (B), and within the periphery (C). Dotplots show association between relative rich-club connectivity and (D) T25FW = timed 25-foot walk, (E) SDMT = Symbol Digit Modalities Test, (F) NHPT = Nine-Hole Peg Test dominant hand, and (G) NHPT nondominant hand. Lines and colored areas represent regression estimates and their confident intervals. For details, see Methods and Results sections.

Absolute rich-club connectivity as well as proportional distribution of connections between the rich club and periphery was not associated with sex, age, disease duration, or T1 lesion volume. By contrast, relatively lower rich-club connectivity and higher peripheral connectivity were associated with increasing T2 lesion load (τ = −0.21, p = 0.034) and lower gray matter volumes (τ = 0.21, p = 0.037). Relatively lower rich-club connectivity was associated with NHPT (nondominant: τ = −0.26, p = 0.014) and T25FW (τ = −0.20, p = 0.047). An increased relative peripheral connectivity was instead linked to lower cognitive performance on the SDMT (τ = −0.20, p = 0.049). FDR-corrected p values remained significant for NHPT and relative rich-club connectivity. Post hoc, an adjustment for brain volume and T2 lesions did not improve the predictive value of rich-club connectivity for T25FW performance (p = 0.283). Concerning NHPT and SDMT, the adjustment for brain volume and T2 lesions performed better than the simple models (both p < 0.001). Alternative rich-club definitions did not show a better ability to discriminate between patients with PPMS and HCs, nor were they more closely associated with disability (table e-1).

DISCUSSION

The rich-club organization of the human connectome has been identified as a fundamental feature of brain networks.[6] Here, we investigated for the first time how the rich-club organization of the human structural connectome is affected in patients with PPMS. Generally, we observed a preserved rich-club organization in patients with PPMS. However, compared with HCs, the connectivity within the rich club was reduced. In addition, lower rich-club connectivity was associated with higher disability.

Based on the weighted rich-club effect, we confirmed the rich-club organization in controls and patients. The nodes forming the rich club in controls did not differ from those in moderately disabled patients with PPMS. Although the global connectivity was lower in patients, the fundamental organizational principles did not seem to be affected. Superior frontal, precuneus, superior parietal, and insular cortex in both hemispheres formed the rich club in patients and controls in line with the originally defined rich-club nodes.[6] The dominance of these nodes was consistent across different nodal graph metrics.

The absolute number of rich-club connections was lower in patients than that in controls, whereas peripheral and feeder connections did not differ. Adjusted for the total connectivity of individual brains, we observed an altered distribution of connections between the rich club and periphery. Specifically, in patients with PPMS, we observed a lower density of rich-club connections compared with controls, whereas peripheral connectivity was relatively increased, which might be due to a lower extent of connectivity loss in the periphery compared with the rich-club or might result from compensatory peripheral rewiring. Our adjustment strategy demonstrates that the observed loss of rich-club connections cannot be explained solely by the global loss of connectivity but indicates a disease-specific pattern. Here, reduced rich-club connectivity was associated with T2 lesion load and gray matter volume, connecting our findings to previous observations.[15] Moreover, we observed a correlation of the change in rich-club connectivity with clinical measures of mobility, hand function, and cognition in our patients.

Alterations to the distribution of connections between rich-club and peripheral nodes have also been observed in other neuropsychiatric diseases and support a disease specificity of distinctive patterns. In migraine patients, an increased number of feeder connections is associated with a higher network efficiency and suspected to cause a higher integration of subnetworks involved in pain processing.[10] However, patients with schizophrenia show reduced rich-club connectivity,[25] and cognitive decline and reduced general function are associated with a pronounced loss of rich-club connections over 3 years.[9] Predominantly neurodegenerative diseases such as Parkinson disease seem to be associated with a pronounced loss of peripheral and feeder connections.[11,26] Within this context, our findings indicate that diffuse perturbations of brain connectivity in PPMS result in a pronounced loss of connectivity within the rich club, and that the overall pattern of connectivity loss differs from primarily neurodegenerative diseases. However, this cross-sectional study does not allow to distinguish between the impact of inflammatory and neurodegenerative mechanisms on the assessed connectivity loss.

This interpretation is supported by data indicating that atrophy appears to follow nonrandom patterns in MS correlated with cognitive impairment and disability.[12] Most of the patterns include the insula, which belongs to the rich-club regions. Moreover, a close correlation between T2 lesions and reduced local efficiency of the insula has been described before,[15] and the affection of hub regions in early RRMS has recently been confirmed in structural and functional connectomes.[27] These findings support our observation that PPMS affects particularly the regions of the brain with a prominent and integrative role in the structural connectome, whereas peripheral connections are less compromised, preserved, or even rewired due to neurorepair. A longitudinal study of fMRI connectivity found a compensatory upregulation in early disease stages without relevant disability.[28] A global loss of connectivity was observed in more disabled patients and correlated with accumulation of disability. Cross-sectional studies support the observation that functional reorganization in MS is complex and leads to different patterns of activation and deactivation related to disability.[29] However, the benefit of this reorganization is questionable.[30] Recruitment of additional neural resource might be inefficient and rather maladaptive than adaptive.[30]

The topology of network reorganization can as well be deduced by contrasting alterations in the large-scale functional networks.[31] The default mode network is considered to represent a backbone of structural-functional organization and shows a heterogeneous pattern of activation and deactivation in MS.[32] Altered default mode connectivity contrasts the connectivity in more peripheral networks, adding further evidence for inverse effects in hubs and the periphery.[31] However, it remains an unresolved issue how reorganization of functional connectivity needs to be interpreted and what the underlying pathologic or repair mechanisms are. Moreover, it is unknown how these functional changes translate into altered structural connectivity. The cross-sectional nature of our study does not allow determining when the shift from rich-club to peripheral connectivity occurs. However, preliminary data in RRMS indicate that the shift toward more peripheral connectivity is detectable within a time frame of 6 months and might be reduced by an exercise intervention.[33] A cross-sectional study in RRMS found also an increased structural connectivity on a local level.[34] Taking the functional data into account, one might hypothesize that a continuous loss of rich-club connectivity can only be compensated by peripheral connections in the early phase.

In contrast to our rich-club analyses, conventional global graph metrics did not differ between patients and controls. The insensitivity of global graph metrics has been described before. For example, they were not able to separate comatose patients from HCs in an fMRI study, while hub regions showed a clear loss in connectivity.[35] These findings support our interpretation that PPMS does not lead to major topological changes but rather to subtle alterations of the brain network.

Our findings are of explorative nature and need further validation. The sample size is relatively small, but in line with similar studies investigating rich-club connectivity.[10,11,26] The study was sufficiently powered to detect alterations in the rich-club connectivity, but was too small to investigate the association with other MRI metrics in depth. For example, we did not detect a difference in white matter volume between patients and controls. Moreover, we could not apply model selection in multivariate statistics to determine the specificity of rich-club alterations for disability in comparison with other MRI measurements. Concerning the mechanisms behind our observations, our interpretations remain hypothetical, as we did not investigate longitudinal changes. Not including relapsing-remitting patients limits the generalization of our findings, but we were primarily interested in neurodegenerative aspects of MS and thus aimed to restrict the influence of acute inflammatory disease activity as much as possible by studying a PPMS cohort. However, especially spinal cord pathology in this patient group might influence disability by means of the EDSS and mobility assessment and make the interpretation of our association results more difficult. Furthermore, the applied method to reconstruct connectomes shares some inherent limitations. Although the method is well accepted, it may be affected by methodological and disease-specific aspects, such as the undistinguishable, but heterogenic histopathology of MS lesions.[36,37] MS lesions affect the tractography results such as they may underestimate true structural connectivity.[38] However, false-negative connections are known to have a lower impact on the network topology than false-positive connections,[36] and in our data set, the global connectivity did not differ between patients and controls. Our analysis followed the original analysis of the rich club that restricted the topological analyses to cortical regions.[6,25] Thus, we cannot conclude about the role of subcortical gray matter or the cerebellum in the rich-club topology.

In conclusion, PPMS seems to induce a loss of structural connections between important brain regions. Our findings emphasize to interpret the impairment of rich-club connections as disease specific and disability related. Thus, rich-club connectivity is a promising indicator for understanding and monitoring the patterns and mechanisms of neurodegeneration in PPMS.